How should you report a study with human and synthetic responses?

Start with the source of each result

A buyer reviewing a boosted study needs to know which findings came from people and which came from simulation. The AAPOR task force's 2026 report on responsible AI integration calls for disclosure of AI's role, human oversight or validation, and the number of human respondents. It says AI-generated cases are not research participants.

The report also cautions that counts of AI instances can be ambiguous and are not a useful quality measure. If your delivery includes generated-row counts, use them to account for the files. Do not present those counts as additional people, an effective sample size, or evidence of accuracy.

The checklist below is our proposed buyer reporting format. It extends the source-labeling principle into a client presentation; it is not an AAPOR certification or a guarantee that blending is appropriate.

Label the chart, not only the methods appendix

Use labels such as 'Human survey result,' 'Synthetic output,' and 'Blended estimate' wherever each figure appears. A reader should be able to identify the source without searching another document. State the target audience and the base for that question, including any exclusions.

AI can also interview people or code their answers. Those uses do not turn the human respondents into synthetic respondents. Describe the actual role: who supplied the answers, what the AI changed or generated, and what a researcher reviewed.

For a blended figure, show the human result alongside it and identify the combination rule. If the method changes the influence of the two sources, explain that influence rather than relying on one combined row count.

Illustrative example: the blend changes the headline

Illustrative example. All counts below are invented. They describe the same single-choice Yes/No question for the same target audience. Each record has a valid answer and equal weight within its source; there are no missing answers. These are not customer data or Lewsearch performance results.

Simply pooling every row with equal weight gives (60 + 450) / (100 + 900) = 51% Yes. The human records account for 10% of the total weight. Giving the two source estimates equal weight instead gives (60% + 50%) / 2 = 55%. The observed human result remains 60% in both calculations.

Both calculations are arithmetically correct for their stated rules. Neither is a recommendation to blend, and neither demonstrates that the resulting estimate is more accurate. The difference shows why the combination rule belongs beside the result. In an actual study, justify the estimator for the intended population and decision before using it.

A transparent chart note for the first calculation would be: 'Illustrative blended estimate: 51% Yes. Based on 100 human responses and 900 synthetic rows, pooled with equal weight per record; the human records carry 10% of the weight. Human-only result: 60%. No population uncertainty interval is asserted in this example.' Avoid a headline that describes this as 1,000 people surveyed.

Keep uncertainty tied to the estimator

AAPOR's disclosure standards distinguish probability-sample sampling error from precision measures for nonprobability research. For the latter, the standards require a defined measure, the underlying model, validation of its assumptions, and an explanation of the calculation. They also call for disclosure of weighting and sample sources.

The AI task force report's synthetic-response example warns against producing a margin of error from synthetic responses. A larger generated file does not, by itself, justify a smaller population margin of error.

For a combined estimate, ask the analyst to explain what uncertainty measure, if any, is justified by the actual method. State its assumptions and limitations. A narrow spread across repeated synthetic runs describes repeatability under that setup; it does not by itself quantify error against the human population.

Copy this into the report handoff

• Sources and roles: identify human data collection, generated responses, and any AI interviewing or coding. Name the responsible research provider and the review performed.
• Human evidence: show recruitment, field dates, question wording, answer options, unweighted human bases and weighting. Show the base for each reported subgroup, not only the study total.
• Synthetic delivery: label generated outputs, record the method/version identifier and run date, and account for delivered rows and exclusions without presenting them as independent people.
• Combination rule: identify every blended figure, its target, the formula or estimator, and the influence of each source. Preserve the human-only result for comparison.
• Validation: distinguish studies used to calibrate the method from untouched human evidence used to evaluate it. Report relevant misses and scope limits alongside the result.
• Uncertainty and use: explain any interval and its assumptions, say when uncertainty has not been established, and name the decisions that still require additional human evidence.

Ask for the same separation from Lewsearch

Lewsearch's sample-boosting page states that the human n remains the human n. Synthetic rows are separately labeled unless blending is agreed in writing, and blended rows still do not become additional human completes. Use the reporting checklist when agreeing on the deliverable.

For the evaluation itself, use the blind-validation buyer guide. It covers protecting the human holdout and freezing scoring before results are revealed. The reporting format here helps readers interpret a delivery; it does not replace that evaluation.

FAQ

Can I report one total sample size for human and synthetic records?

Keep the human respondent count separate. You can account for the generated file with a clearly labeled row count, but a combined record total must not imply that all records are independent people or establish an effective sample size.

Must every boosted study publish a blended estimate?

No. Human results and synthetic outputs can be delivered side by side. A blended estimate needs a stated purpose, a justified combination method, and clear labeling. The numerical example is not an endorsement of either blending rule.

Does a larger synthetic file reduce the human survey margin of error?

Generating more rows does not increase the number of observed human responses. Any precision claim for a combined estimate must be justified by its method and assumptions; it cannot be inferred from the combined row count alone.